From 044d2f744be1d5910eba5c0fd97b0450d5ecc003 Mon Sep 17 00:00:00 2001 From: danielhanchen Date: Mon, 31 Aug 2026 09:28:13 +0000 Subject: [PATCH] unsloth: pin #149 so GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disables it ggml tests getenv(...) != nullptr, so =0 enables managed allocation. With #157 pinned this is no longer a correctness bug, but it still costs about 15% throughput and can lower the ceiling: managed draws host RAM instead of the device carve-out rather than adding to it, so a model that loads without the variable can be OOM-killed with it. Verified locally to co-merge onto b10705 with #157 and #158, net +79/-5 across 4 files. --- scripts/unsloth/pr-set.json | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/scripts/unsloth/pr-set.json b/scripts/unsloth/pr-set.json index fe511789313..3e69aa63d3d 100644 --- a/scripts/unsloth/pr-set.json +++ b/scripts/unsloth/pr-set.json @@ -26,6 +26,7 @@ "https://github.com/ggml-org/llama.cpp/pull/27754/commits/5796547f37f5943513dfa130065ec88f9e30e0f5", "https://github.com/unslothai/llama.cpp/pull/137/commits/4e1865e34ec5f6ca39403215c89129c13731be70", "https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341", - "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b" + "https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b", + "https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733" ] }